Graham,
The problem on Win32 is that (I believe) we never want to initialize Python
in the persistent parent process. All the web action is in the child
process which is long-lived and it is this child process that maintains the
thread pool which services web requests.
The parent process as far as I can tell sits there to support restarting the
child process and support the Win32 Service Control Manager (SCM) which has
a protocol for how a process must respond to certain messages in order to be
a service on Win32.
I do not know how to use flags alone to distinguish the two processes. The
code I put in is not trying to restrict a call to python_init() to only
happen once in the parent process. It is preventing python_init() from
initializing Python in the parent process.
I hope this clarifies things somewhat.
I want to note here that mpm_winnt.c line 1108 looks like this:
/* AP_PARENT_PID is only valid in the child */
pid = getenv("AP_PARENT_PID");
if (pid)
{
/* This is the child */
This environmental is how it knows to run certain code only in the child
process.
In summary,
if we do not want to run python_init() in the special Win32 parent process,
we need a way to distinguish this parent process from the child process in
which we DO want to run python_init(). The code which maintains this dual
process architecture (undoubtedly in support of the Win32 service
architecture) uses an environmental that it purposefull creates and injects
into the child process "AP_PARENT_PID". I don't see how we can do better
than use this same distinguishing characterisic to know not to run
python_init() in the parent process.
----- Original Message -----
From: "Graham Dumpleton" <[EMAIL PROTECTED]>
To: "Jeff Robbins" <[EMAIL PROTECTED]>
Cc: "python-dev list" <python-dev@httpd.apache.org>
Sent: Tuesday, November 07, 2006 06:02
Subject: Re: MODPYTHON-195
On 04/11/2006, at 12:34 PM, Jeff Robbins wrote:
Graham,
I haven't had any new ideas about this problem. It is clear that on
Windows, mod_python is initialized both in a parent process and more
usefully in the child process that spins up the threads that service
client requests. The parent process is long-lived and the standard hack
to wait for the second call to the ap_hook_post_config is useless
because each "restart" of apache is yet another call (third, fourth,
fifth, etc...) and each time there's a leak of one handle.
The fix I tested seems reasonable. I know it is dependent on
mpm_winnt.c, but, after all, that file is the file responsible for the
dual process architecture on windows to begin with. And the fix has an
#ifdef win32 so it won't hurt linux users.
I'd like you to consider folding it in. I think it is better than
having a leak (along with spurious python initialization) on windows.
Jeff, can you see if you can come up with a test based on 'initialized'
and
'child_init_pool' as I note in:
http://issues.apache.org/jira/browse/MODPYTHON-195
If it is only in the parent process you need to skip subsequent calls,
perhaps:
if (child_init_pool == 0 && initialized != 0)
return OK;
Will have to think about how this may screw up old versions of Mac OS X
though which is why initialized was added in the first place.
You might include in your debug a call to Py_IsInitialized() so it can be
determined if Python thinks it is already initialised. Since your fiddle
is
working, I'd say it probably is.
Also see if main_server is set and not zero as well and whether its value
is different to 's' passed as argument to function. Whether it is the
same
or not may dictate where in function the check to bail out of function
needs
to be. It may have to go just before the global config and mutexes are
created.
Graham
